ROSSMANN Stores Sales Predictions

Summary

1. Context
2. Challenge
3. Solution Development
4. Conclusion and Demonstration
5. Next Steps

1. Context

• Monthly Results Meeting
• CFO asked for a Sales Forecast for the Next 6 Weeks for each Store

2. Challenge

Problem

• Budget definition for store renovation

Causes

• Current sales forecasts were highly inconsistent.

• The sales forecast process is based on past experiences.

• All sales forecasts are made manually by Rossmann's 1,115 stores.

• Sales viewing is limited to computers.

Solution

• Using Machine Learning to forecast sales for all stores

• Sales predictions can be viewed on a smartphone.

3. Solution Development

Data Description

Data Dimension

Number of Rows 1017209
Number of Cols 18

Descriptive Statistics

	attributes	min	max	range	mean	median	std	skew	kurtosis
0	store	1.0	1115.0	1114.0	558.429727	558.0	321.908493	-0.000955	-1.200524
1	day_of_week	1.0	7.0	6.0	3.998341	4.0	1.997390	0.001593	-1.246873
2	sales	0.0	41551.0	41551.0	5773.818972	5744.0	3849.924283	0.641460	1.778375
3	customers	0.0	7388.0	7388.0	633.145946	609.0	464.411506	1.598650	7.091773
4	open	0.0	1.0	1.0	0.830107	1.0	0.375539	-1.758045	1.090723
5	promo	0.0	1.0	1.0	0.381515	0.0	0.485758	0.487838	-1.762018
6	school_holiday	0.0	1.0	1.0	0.178647	0.0	0.383056	1.677842	0.815154
7	competition_distance	20.0	200000.0	199980.0	5935.442677	2330.0	12547.646829	10.242344	147.789712
8	competition_open_since_month	1.0	12.0	11.0	6.786849	7.0	3.311085	-0.042076	-1.232607
9	competition_open_since_year	1900.0	2015.0	115.0	2010.324840	2012.0	5.515591	-7.235657	124.071304
10	promo2	0.0	1.0	1.0	0.500564	1.0	0.500000	-0.002255	-1.999999
11	promo2_since_week	1.0	52.0	51.0	23.619033	22.0	14.310057	0.178723	-1.184046
12	promo2_since_year	2009.0	2015.0	6.0	2012.793297	2013.0	1.662657	-0.784436	-0.210075
13	is_promo	0.0	1.0	1.0	0.155231	0.0	0.362124	1.904152	1.625796

• Rossmann was founded in 1972. Values of Competition_open_since_year lower than 1972 indicate the years when closest competitors, from other pharmacy chains, were opened.

• The competition_distance variable has high positive values of skew and kurtosis, indicating that the distribution is skewed right and has a large tail.

• There is a vast difference in the range in some features. Higher ranging numbers have superiority of some sort. So these more significant number starts playing a more decisive role while training some models. It's needed to apply some sort of scaling over the features.

Mind Map Hypothesis

Exploratory Analysis Hypotheses

H1. On average, stores with a larger assortment should sell more.

H2. Stores with closer competitors should sell less.

H3. Stores with competitors that have been around for longer should sell more.

H4. Stores with more consecutive promotions should sell more than stores with regular promotion

H5. Stores open during the Christmas holidays should sell more.

H6. Stores should sell more over the years.

H7. Stores should sell more in the second half of the year

H8. Stores should sell less after the 10th of each month.

H9. Stores should sell more on average on weekends.

H10. Stores should sell less during school holidays.

Exploratory Data Analysis

Response Variable

Numerical Variables

/tmp/ipykernel_50303/2232118446.py:6: UserWarning: To output multiple subplots, the figure containing the passed axes is being cleared.
  num_attributes.hist(bins=25, ax=ax);

Categorical Variables

Hypothesis Validation

H1. On average, stores with a larger assortment should sell more.

True stores with a larger assortment sell more on average

H4. Stores with more consecutive promotions should sell more than stores with regular promotion

Falso Stores with more consecutive promotions sell less

H9. Stores should sell more on average on weekends.

False, there is not enough evidence to conclude that sales on weekends are greater than sales on weekdays.

Summary of Hypotheses

Hypotheses    Conclusion    Relevance
------------  ------------  -----------
H1            True          High
H2            False         Low
H3            False         Low
H4            False         High
H5            True          High
H6            True          Low
H7            True          High
H8            True          High
H9            False         High
H10           False         High

Multivariate Analysis

Numerical Attributes

Categorical Attributes

Machine Learning Modelling

Compare Model's Performance

	Model Name	MAE	MAPE	RMSE
0	XGB Regressor	694.066535	0.102831	999.914943
1	Random Forest Regressor	747.458229	0.111702	1098.595402
2	Average Model	1429.763326	0.216814	1939.328730
3	Linear Regression	1867.623495	0.296267	2657.022835
4	Linear Regression - Lasso	2192.664126	0.343490	3092.842416

4. Conclusion and Demonstration

Business Performance

	store	predictions	worst_scenario	best_scenario	MAE	MAPE
291	292	108,383.78	105,018.11	111,749.45	3,365.67	0.61
908	909	244,502.50	237,037.94	251,967.06	7,464.56	0.52
875	876	200,492.30	196,658.64	204,325.96	3,833.66	0.29
721	722	356,656.97	354,564.33	358,749.61	2,092.64	0.28
594	595	383,771.44	379,840.58	387,702.29	3,930.85	0.27
...	...	...	...	...	...	...
493	494	326,901.41	326,470.16	327,332.65	431.25	0.06
373	374	258,871.70	258,486.16	259,257.24	385.54	0.06
561	562	747,947.50	746,976.29	748,918.71	971.21	0.06
958	959	255,862.62	255,458.45	256,266.80	404.17	0.05
259	260	232,898.20	232,580.29	233,216.12	317.91	0.05

1115 rows × 6 columns

Total Performance

	Scenario	Values
0	predictions	US$290,943,413.95
1	worst_scenario	US$290,165,597.86
2	best_scenario	US$291,721,230.03

Machine Learning Performance

Demonstration - Telegram

5. Next Steps

• Model Workshop for Business Users
• Collect Usability Feedback
• Improve model performance (MAPE) by 5%

Q & A

Thank You Very Much!